ModuleOp module = getOperation();
    if (!HasTfSavedModelSemantics(module)) {
      return;
    }

    auto init_op = GetSessionInitializerOp(module);
    if (init_op == nullptr || init_op.getInitializers().empty()) {
      return;
    }

    SymbolTable symbol_table(module);
    for (auto initializer : init_op.getInitializers()) {
      auto func = symbol_table.lookup<func::FuncOp>(

    ],
    licenses = ["notice"],
)

package_group(
    name = "friends",
    packages = [
        "//tensorflow/compiler/mlir/tfr/...",
    ],
)

gen_op_libraries(
    name = "mnist_ops",
    src = "ops_defs.py",
    deps = [
        "//tensorflow:tensorflow_py",
        "//tensorflow/python/framework:ops",
        "//tensorflow/python/ops:math_ops",
        "//tensorflow/python/ops:math_ops_gen",

		fmt.Printf("words = %3d deltaT = %10s (%4d%%) is %s better: %v", i, delta, percent, upper, pos)
		if i == from {
			initPos = pos
		}
		if threshold == 0 && pos != initPos {
			threshold = i
			fmt.Printf("  threshold  found")
		}
		fmt.Println()

	}
	if threshold != 0 {
		fmt.Printf("Found threshold = %d between %d - %d\n", threshold, from, to)
	} else {

# CHECK: "tf_saved_model.global_tensor"()
# CHECK: "tf_saved_model.session_initializer"() <{initializers = [@[[init:.*]]]}> : () -> ()

# CHECK:      func @[[init]]
# CHECK-SAME: tf_saved_model.initializer_type = "init_op"
# CHECK-NEXT: [[R6:%.*]] = "tf.Const"()
# CHECK-NEXT: [[R5:%.*]] = "tf.Const"() <{value = dense<[1, 2,
# CHECK-NEXT: [[R7:%.*]] = "tf.HashTableV2"()
# CHECK-SAME: shared_name = "[[hash_table:.*]]"

namespace tensorflow {
namespace internal {

// A SavedModel may store the name of the initialization op to run in the
// in the SignatureDef (v2) or a collection (v1). If an init_op collection
// exists, then the collection must contain exactly one op.
Status GetInitOp(const string& export_dir, const MetaGraphDef& meta_graph_def,
                 string* init_op_name) {

# CHECK-SAME: [[ARG0:%.*]]: tensor<!tf_type.string> {tf_saved_model.bound_input = @[[asset0]]}
# CHECK-SAME: [[ARG1:%.*]]: tensor<!tf_type.string> {tf_saved_model.bound_input = @[[asset1]]}
# CHECK-SAME: tf_saved_model.initializer_type = "init_op"
# CHECK-NEXT: [[R0:%.*]] = "tf.HashTableV2"()
# CHECK-SAME: shared_name = "[[hash_table:.*]]"
# CHECK-NEXT: "tf.InitializeTableFromTextFileV2"([[R0]], [[ARG0]])


def write_vocabulary_file(vocabulary):

# CHECK-SAME: tf_saved_model.exported_names = ["__tf_saved_model_session_initializer_init"]
# CHECK-SAME: tf_saved_model.initializer_type = "init_op"
# CHECK-SAME: }
# CHECK-DAG: %[[CST_0:.*]] = "tf.Const"() <{{{.*dense<.*> : tensor<2xi32>.*}}}> {{{.*}}} : () -> tensor<2xi32>
# CHECK: %[[RAND_STD_NORMAL:.*]] = "tf.RandomStandardNormal"(%[[CST_0]])

  // CHECK: "tf_saved_model.session_initializer"()
  // CHECK-SAME: initializers = [@init_func_init_op]

  func.func @init_func_init_op() -> () attributes {
      tf_saved_model.initializer_type = "init_op",
      tf_saved_model.exported_names = ["tf_saved_model.session_initializer_init_op"]} {
    %cst_0 = "tf.Const"() {value = dense<1.000000e+00> : tensor<2xf32>} : () -> tensor<2xf32>

inline constexpr StringRef kTfSavedModelInitializerRestoreType = "restore_op";

// Indicates that the initializer corresponds to the init op.
inline constexpr StringRef kTfSavedModelInitializerInitType = "init_op";

class TensorFlowSavedModelDialect : public Dialect {
 public:
  explicit TensorFlowSavedModelDialect(MLIRContext *context);
  LogicalResult verifyRegionArgAttribute(Operation *op, unsigned region_index,